EU AI Act · Article 50 · transparency
The transparency cliff.
If your AI talks to people or generates content, Article 50 lands on you. Disclosure and machine-readable marking of AI content — with real fines (up to €15M or 3% of turnover) and enforcement from day one.
click to launch eu
Measurement Lens
417 provisions · 4 axes · deterministic
The full checklist — and what we do about each line
We sell Article 50 tooling, so each obligation below is paired with what happens on this site — not what a brochure says should happen.
If people interact with an AI system, they must be told it is an AI system at the time of the first interaction — in a way that is clear and distinguishable from the rest of the page, and that meets the applicable accessibility requirements. The only escape is when it is obvious from the circumstances.
On this site: Every surface on this site is classified in a public registry. The ones that talk to a live model mount a first-interaction notice above the input — a keyboard-reachable live region that collapses to a persistent marker, never disappears. The ones that are deterministic say so instead of staying vague.
The registry, surface by surface →Providers of AI systems that generate synthetic audio, image, video or text must mark the output as artificially generated in a machine-readable way — detectable, effective, interoperable, robust. This is the direction the C2PA content-provenance ecosystem and the Code of Practice on marking are converging on.
On this site: Nothing on this site generates synthetic audio, image, video or text, so there is no output to mark — the registry says so out loud. We also measured whether the ecosystem's existing marks survive in the wild: 0 of 20 assets kept their provenance. If we ever ship generative output, the marking obligation attaches immediately — no grace period for features launched after 2 Aug 2026.
The 0-of-20 provenance finding →People exposed to an emotion-recognition or biometric-categorisation system must be informed that such a system is operating on them, and it must process their data under the data-protection rules.
On this site: This site operates no emotion-recognition and no biometric-categorisation system. Nothing here infers how you feel or what you are.
AI-generated or -manipulated image, audio or video that resembles real people, places or events (a deepfake) must be visibly labelled as artificial. Text generated to inform the public on matters of public interest must be disclosed as AI-generated — with one exception: when a natural or legal person holds editorial responsibility and the text has been through human review, the disclosure duty does not apply.
On this site: No generated media is published here. The words on this site are written by people; where tooling assists research or drafting, a named person reviews and holds editorial responsibility before anything ships — which is exactly the discipline the exception is designed to reward, not a loophole to hide behind.
The disclosure must meet the accessibility requirements of the European Accessibility Act (Directive (EU) 2019/882), and it must reach the person — which in practice means clear language the user can understand.
On this site: The notice is a real live region with role and tab order, not a tooltip; dismissing it leaves a permanent one-line marker rather than removing it. The Sovereign answers in the visitor's own language where it can, keeping statutory names in their canonical form.
How the notice is built →Article 50 does not spell out an evidence duty, but enforcement will ask one question: prove the disclosure was in place on a given date. A claim without an artefact is half a compliance story.
On this site: The surface registry lives in source control, every change to it is a commit, and a guard in the release gate fails the build if a surface starts calling a model without being registered and noticed. The public decision chain is recomputable in your browser.
Recompute the chain →The timeline
We sell this. Here is us, under it.
From 2 August 2026 this page's obligations apply to csoai.org itself. So the whole site is measured the same way we would measure yours: 170 interactive surfaces classified — 128 rule-based, 42 AI-system routes across 14 components — each with its mechanism and its notice state published. Where a surface is not yet classified it defaults to the strictest reading, not the friendliest.
Surface classification, first-interaction disclosure patterns, C2PA / content-provenance marking, and Layer-0 signed evidence that it was in place — reproducible for auditors.
Frequently asked
What does EU AI Act Article 50 require?
Article 50 sets transparency duties: AI systems that interact with people must disclose they are AI at the first interaction; providers of generative AI must machine-readable-mark synthetic audio, image, video and text as AI-generated; people exposed to emotion-recognition or biometric-categorisation systems must be told; and deepfakes plus AI-generated public-interest text must be clearly labelled — the last with an exception where a person holds editorial responsibility after human review.
When does Article 50 take effect?
Article 50 transparency obligations — and the enforcement powers behind them — apply from 2 August 2026. Systems already on the market have a marking grace period until 2 December 2026 for synthetic content, and GPAI models placed on the market before 2 August 2025 have until 2 August 2027. Anything launched after 2 August 2026 must comply immediately.
What are the penalties for breaching Article 50?
Non-compliance with transparency obligations can attract fines of up to €15 million or 3% of total worldwide annual turnover, whichever is higher, alongside national market-surveillance enforcement. (Higher tiers exist for other articles — up to €35M or 7% for prohibited practices.)
What is the editorial-responsibility exception?
AI-generated text published to inform the public does not have to be labelled as AI-generated if a natural or legal person holds editorial responsibility for it and it has undergone human review. It rewards genuine editorial control; it is not a way to launder unreviewed generated text through a nominal editor.
How does CSOAI help with Article 50 compliance?
CSOAI maps Article 50 to concrete, verifiable controls: a surface registry classifying every interactive surface, first-interaction disclosure patterns, C2PA / content-provenance watermarking for machine-readable marking, and Layer-0 (Ed25519) signed evidence that the marking and disclosure were in place — reproducible for auditors. And it applies the same controls to itself: every surface of this site is classified publicly at /ai-transparency.
Our own Article 50 self-conformance
source: ai-surfaces registry v2.0.0 + /ai-transparency record, 2026-08-01
Frequently asked questions
Article 50 transparency duties, in plain terms — and what we do on this site ourselves.
What does EU AI Act Article 50 require?
Providers and deployers must disclose when a person is interacting with an AI system, label AI-generated content, and disclose emotion-recognition or biometric categorisation where used. The main transparency obligations apply from 2 August 2026.
Does this site comply with its own reading of Article 50?
We publish our self-conformance record openly: 138 routes classified in our registry (38 AI-system surfaces), notices mounted at first interaction, and a public corrections history — including where our own first draft got it wrong.
Is an Article 50 notice the same as conformity?
No. A notice satisfies a transparency duty; it is not a conformity assessment and does not certify anything. Our record states what was measured, when, and what remains open.
How do I check whether a page I am on uses AI?
Every AI-system surface on this site carries a notice at first interaction, and the full classification is published at /ai-transparency — dated, hashed, and updated when routes change.
Dates verified July 2026 (AI Omnibus provisional agreement, May 2026). Indicative guidance, not legal advice — verify against primary EU sources.